Did you know that you can navigate the posts by swiping left and right?

Pay the loss once : kilohertz entanglement between an ion and a diamond

15 Oct 2026 . category: networking . Comments
#networking #quantum #datacenter

two posts on this blog ended on the same complaint from opposite ends.

the store-and-forward post was about a silicon vacancy in a diamond cavity, a memory that interacts with almost every photon that reaches it. the experiment produced about a secret bit a minute, and the bottleneck was on the other side of the link: nothing was feeding the memory photons quickly enough. the long-lived ion post was about trapped calcium ions, which emit a photon whenever you ask, and that link heralded about twice a second because only 2.6% of the photons made it into a fiber.

one platform is good at catching photons and the other is good at making them. this paper connects the two.

Hartung et al., “Quantum Networking at the Speed of Quantum Computation: Kilohertz Entanglement in a Heterogeneous Quantum System,” arXiv:2610.10705, posted on 7 October by IonQ with Duke. a barium ion in a room-temperature trap, a silicon vacancy in a diamond nanocavity at 100 mK in a dilution refrigerator, and between them one photon that changes color, changes encoding, bounces off the diamond, changes encoding back and lands on a detector. one click means the ion and the diamond now share an entangled pair. 1,032 pairs a second at 87.9% fidelity, four times the rate of the fastest photonic link anyone has built between two ions.

the author list reads like a merger, because it is one. on one side is the Duke ion group whose 2024 link holds that ion-ion record. on the other are the Harvard and AWS diamond people who founded Lightsynq, which IonQ announced it was buying in May 2025. this is the paper where the two halves get plugged into each other.

the trick is in one word up there: one click. below: why a link between two ions pays its loss twice, how this one pays it once, where 29.7 dB goes, why the attempt clock runs at 1.1 MHz and what would make it faster, the knob that trades rate for fidelity, and what happens to the four-fold advantage when you ask for the kind of link a quantum computer needs.

the pitch for modular quantum computing is the pitch for a data center. nobody builds one enormous server. you build racks of ordinary ones and put a network between them. the data center post covered the topology half of that problem. this paper is about the link itself.

the link’s job is to hand out entangled pairs between modules, and a two-qubit gate across the boundary uses up one pair each time it runs. the estimates the paper cites for running logical operations continuously between error-corrected modules land between 10 and 100 kHz, with each pair allowed about ten times the error of a local gate. an MIT and Lincoln Lab group put finer numbers on it this summer (Knollmann et al.). for one inter-module gate per error-correction cycle on a large trapped-ion machine, they count 1.5 × 10³ pairs per second per module with code surgery at a slow 55 ms syndrome cycle, and 8 × 10⁴ at a fast 1 ms one, all at 99.9% fidelity.

the best photonic link between two trapped ions delivers 250 pairs per second (O’Reilly et al., PRL 2024). at the start of this paper, then, the rate gap is somewhere between a factor of six and a factor of three hundred depending on whose numbers you use, and nobody has mentioned fidelity yet.

paying twice

here is how that record link works. each of the two ions is excited, and each decays and emits one photon whose polarization is entangled with its spin. both photons travel to a beam splitter in the middle. if the two detectors behind it click in the same attempt, the photons interfered in a way that only happens when the two ions are left in a Bell state, and the attempt has succeeded.

both is the expensive word. in the Duke experiment each ion got a photon onto a detector 2.2 to 2.3% of the time. two of those in the same attempt is 0.05%, and the beam splitter only recognizes two of the four Bell states, which halves it again: 2.5 × 10⁻⁴ per attempt. at a million attempts a second that is 250 pairs.

in link budget terms each ion side is a 16.5 dB budget, and a two-photon herald pays it twice, plus 3 dB for the Bell states it can’t see:

two-photon herald:   2 × 16.5 dB  +  3 dB   ≈  36 dB   →  1 success in 4,000 attempts

this is the coincidence tax again, at the scale of a room instead of 350 km: two independent, unreliable events have to land in the same slot, so their probabilities multiply.

there are two ways to stop paying twice. one is single-photon heralding: excite both ions weakly, interfere the two possible photons, and accept a single click as success. the long-lived ion experiment ran on it. it pays the loss once, but it charges you in fidelity, because the harder you excite the ions the more often both emit and spoil the state, and it turns the two fibers into an interferometer whose phase has to hold to a fraction of a wavelength.

the other is what this paper does. make only one photon, and instead of sending it to meet a second photon at a beam splitter, have it meet a memory.

six hops for one photon

the photon makes one round trip and changes form five times along the way. the widget follows a single attempt through the six steps of the paper’s Fig. 1.

one attempt, followed to the clickthe six steps of the paper's Fig. 1. pick a step to see what the photon carries, the joint state at that point, and how many of the roughly 774,000 attempts made each second still have a photon in play.493.5 nm, the ion's color737.45 nm, the diamond's colorion spindiamond spin
step
 
attempts per second that still have a photon in play
still in play-photons per second
lost at this step-
spent so far-of 28.1 dB from attempt to click
what is flying-

the ion picks a polarization

a picosecond pulse at 493.5 nm lifts the ion out of spin up with 97% probability, and within nanoseconds it falls back one of two ways. it can drop to spin down and emit a π-polarized photon, or drop back to spin up and emit a σ⁻-polarized one. the polarization now records which way the spin went, and that shared record is the entanglement. about a quarter of the time the ion falls into a third, metastable level instead, and that attempt is over before it began. a 650 nm repump clears it for the next one.

the objective looks at the ion from the side, and a singlemode fiber maps π onto vertical polarization and σ⁻ onto horizontal. there’s a nice cancellation hiding in the supplement. the π decay is half as likely as σ⁻, but a π photon prefers to leave sideways, straight into the objective, and it also couples into the fiber better (78% against 57% for an ideal lens) because the σ⁻ photon carries orbital angular momentum the fiber mode doesn’t want. the effects net out to 51:49 in the fiber, so the entangled state comes out balanced without anyone having to balance it.

a new color

the silicon vacancy only talks to light at 737.45 nm. a periodically poled lithium niobate waveguide, pumped at 1492.3 nm, subtracts the energy of a pump photon from the ion’s photon and leaves exactly that wavelength. the crystal only converts one polarization, so it sits inside a Sagnac loop: H goes around one way and V the other, both get converted, and they recombine with a relative phase that is stable by construction. 54% of the photons convert and 88% of those get through the filters.

the pump has a side effect. with the ion’s light blocked, the detectors still click at 1/27 of the normal rate. that is noise the 5 W pump generates in and around the signal band, getting past the filters, and every one of those clicks is a fake herald.

polarization into time

the cavity accepts a single polarization, so the qubit has to move out of polarization and into time. an unbalanced interferometer sends V down a short arm and H down about 15 m of fiber, which turns the photon into an early pulse and a late pulse 72 ns apart, and a Pockels cell then rotates the late pulse to match the early one. the spacing isn’t arbitrary. the photon is 48 ns long, and the 18 ns microwave pulse that flips the memory has to fit between the two halves.

the bounce

the store-and-forward post spent a long time on this part, so briefly. at the operating frequency, the cavity reflects light only when the electron spin is in its bright state. the spin starts in a superposition. the early pulse arrives and is reflected only if the spin is bright, a microwave π pulse flips the spin, and then the late pulse gets the same treatment. if anything comes back, the early pulse went with one spin state and the late pulse with the other, and the ion, the photon and the spin are now tied together.

two things fall out of that design. the first is that this is an amplitude gate: the photon only comes back in the branches where it meets the bright state, and on average that is half of them. the cavity can never hand back more than about half the photons it receives, however good the device is, and this one hands back 44%. the second is that the other 56% aren’t harmless. a photon that went into the cavity and never came back still interacted with the spin, so it leaves the memory partly measured, and nobody gets told. that sets the rate-fidelity trade further down.

the way back, and the click

on the return trip the Pockels cell sends each pulse through the other arm of the same interferometer, so both pulses have now traveled the same total distance and come out together, back in polarization. the interferometer’s own phase drops out as well, because every photon samples both arms once. the store-and-forward experiment needed a 142 ns interferometer packed into a sand-filled briefcase and relocked for half of every 400 ms. this one needs no stabilization at all, as long as it holds still for the 300 ns of the round trip.

a measurement at 45° then erases which pulse was which, and a click on either detector projects the ion and the spin onto a Bell state, with the sign set by which detector fired. the controller fixes that sign in real time, along with a phase that depends on exactly when the ion emitted, so every herald delivers the same state. the paper makes a point of this, and it’s right to: a pair whose label you only learn afterwards is fine for a Bell test and useless as the input to a gate.

29.7 dB

here’s the whole link as a budget, from the 1.116 MHz clock to a click. it closes. multiplied out, the paper’s independently measured stage efficiencies predict 1.21 ± 0.20 kHz, and the measured rate is 1.03, inside the error bar. the authors suspect the missing 14% is birefringence in the interferometer. i like a paper that prints its own budget and then checks itself against it.

StageWhat's in itEfficiencyCost
Emissionoptical pumping 0.82, excitation 0.97, decay to the right levels 0.730.582.4 dB
Collection16.9% of the sphere, 85% through the optics, 40% into fiber0.05712.4 dB
Frequency conversionconversion 0.54, filters 0.880.483.2 dB
Basis convertertwo passes at 0.590.354.6 dB
Fridge and cavity couplingfiber to the cryostat 0.90 both ways, cavity coupling 0.89 in and out0.711.5 dB
The bouncespin-dependent reflection0.443.6 dB
Detectionsuperconducting nanowire detectors0.900.5 dB
Protocolshare of the clock not spent resetting or initializing0.691.6 dB
Total1.08 × 10−329.7 dB

a few lines in that table deserve a second look.

half the budget goes on getting the photon into a fiber. 14.8 dB, and 12.4 of it is collection. the objective sees 16.9% of the sphere and the fiber accepts 40% of that. the supplement is frank about why it’s 40% and not the 78% an ideal lens would give: the objective isn’t diffraction limited, with 55% of the light in the central Airy lobe instead of 84%, and an aberration-aware model of this particular lens tops out at 52%.

the interferometer costs as much as the memory. two passes at 59% is 4.6 dB, the same as the entire diamond side (coupling in, the bounce, coupling out). every part in the supplement’s component list is 98% efficient or better, and the supplement blames the difference between the two arms on fiber coupling in the long one. the loss is in the integration rather than the parts, which is exactly the kind of loss integrated photonics is good at removing.

most of the bounce’s 3.6 dB comes from the gate’s design. the cavity does its part well enough. the amplitude gate from above can return about half the photons at best, so roughly 3 of those dB are built into the protocol. the paper’s outlook says a phase gate, which reflects the photon in both spin states, plus better fiber coupling and a better signal chain, should push the link past 5 kHz. from this table that’s plausible. fiber coupling at 70%, the interferometer at 80% a pass, conversion at 70% and a reflection of 0.85 multiply out to about eight times, and to somewhat less once the fixed per-pair housekeeping is counted.

once versus twice

with the budget in hand, the trade fits in a line of decibels. a two-ion link pays the ion’s loss twice plus 3 dB. this link pays it once, plus 13.3 dB of plumbing between the fiber and the click: conversion, two interferometer passes, the fridge, the bounce, the detector. at this ion’s 14.8 dB the memory saves about 5 dB per attempt, which is roughly the factor of four in the headline. but the two lines have different slopes. improve the ion by a dB and the two-photon link gains two, while the memory link gains one.

pay the loss once, or twicechance per attempt of heralding a pair, against how lossy the ion side is (attempt to a photon in fiber). each line is a link budget rather than a fit: two photons pay the ion side twice plus 3 dB, the memory pays it once plus its plumbing, single-photon heralding pays it once plus the price of exciting the ions weakly.two ions, two photonsion and memory, one photontwo ions, one photonmeasured links
ion side
memory side
count
two photons-per attempt, ion-ion
via the memory-per attempt
one photon, two ions-
memory against two photons-

press ion in a cavity and the ranking flips. two photons win by six dB, because once the ion side is cheap, paying it twice costs less than paying 13 dB of plumbing once. the memory wins exactly while the ion is the most expensive part of the budget, and with today’s plumbing the crossover sits at about 9.4 dB per ion side.

two more buttons are worth pressing.

the first is ion-ion, through the memory. this experiment entangles an ion with a diamond. a modular ion computer wants two ions in different modules entangled with each other, and in this architecture that means two ion-diamond links in a row through the same memory: the first parked in the diamond’s ²⁹Si nuclear spin while the second is built, then a measurement on the diamond to swap them. needing two successes in a row costs 3 dB. that moves the crossover to 12.4 dB, about where a diffraction-limited lens would put this ion, and it shrinks the advantage over two photons at today’s ion to 1.7×.

the second is the purple line, single-photon heralding between two ions. at 5% excitation it sits about 3 dB above the memory at every ion efficiency, and it charges for that in fidelity (capped near 95% before any distillation) and in a phase lock across the link. that’s the route the MIT and Lincoln Lab group argues for, combining single-photon heralding, a correction for the ion’s recoil, distillation and photonics built into the trap, and they project 858 pairs per second per channel at 99.9%.

none of this makes the memory a bad idea. it buys two things the decibels don’t show. the photons arriving at the memory never have to be identical to anything, and no fiber ever has to be phase-locked to another. that’s what would let one memory talk to an ion today and to a neutral atom, or a superconducting qubit behind a transducer, later. it’s the premise of DARPA’s heterogeneous architectures program, where IonQ has a contract to build these diamond memories. still, “four times faster” is a statement about this ion and this plumbing, and both of them are moving.

stop and wait

the other half of the rate is how often you get to try. an attempt takes 896 ns, and the paper’s Fig. 2b times it to the nanosecond:

  • 48 ns listening for the previous attempt’s photon
  • 268 ns for the controller to decide whether anything clicked and build the next pulses accordingly
  • 78 ns of cables and electronics
  • 192 ns to put the ion back in its starting state, 160 of them optical pumping, then the excitation pulse
  • 310 ns of photon flight, from the ion through the converters to the diamond and back to the detectors

the ion is busy for less than a quarter of that. the rest of the time the link is waiting to find out whether the last try worked, and it has to wait, because pumping the ion after a successful attempt would erase the pair it just made. this is stop-and-wait ARQ, the protocol everyone learns first, with a window of exactly one.

stop and waitone 896 ns attempt as the paper times it, and what the link does when the controller gets faster, the ion skips optical pumping, the memory moves further away, or several emitters take turns on one collection path.ion busyphoton in flightcontroller and cableslistening for a click
controller
state preparation
one attemptemitters sharing one collection path, one row each
attempt period-
ion busy-of each attempt
pipe holds-emitters before the ion is the bottleneck
fiber loss-there and back, ~3 dB/km at 737 nm
heralds per second-

the decision time is the largest slice, and it’s tangled up with a fidelity problem. after a herald, the next attempt’s pumping light has to be stopped within nanoseconds, so that path runs through two electro-optic modulators that switch in under 100 ns. together they extinguish the light by about 47 dB. what still leaks through after a herald is the largest single ion error in the paper’s budget: 0.87% on XX and YY, 1.45% on ZZ.

drag the distance slider and the other limit shows up. the photon has to reach the diamond and come back before the next attempt can start, so every metre between the ion and its memory adds about 10 ns to every attempt, and at 737 nm the fiber also loses a few dB per kilometre in each direction. at 100 m the rate halves. at a kilometre it’s about 30 pairs a second. one pair per millisecond is a number for modules in the same room, which is what the paper is about. it’s still worth saying out loud, because “quantum networking” in a title reads further than that.

the textbook fix for stop-and-wait is to keep more frames in flight. with the ion idle more than three quarters of the time, about four emitters could take turns on one collection path before the ion became the bottleneck. the diamond side is ready for that: the same team’s chip paper puts 128 cavities on one membrane. the ion side is the hard part, because a high-NA objective only sees a small field, and the paper says so itself.

fifteen tries, then reset

the controller runs attempts in batches of up to 15. if none of them heralds, it re-cools the ion for 3.1 µs, measures and re-prepares the spin, and starts another batch. why not let a batch run to a hundred?

because of the photons that went into the memory and didn’t come back. each attempt has a 0.69% chance of delivering a photon to the cavity, and 80% of those never reach a detector: 56% aren’t reflected, and the return trip loses more than half of what’s left. each of those silent photons is a small, unrecorded measurement of the spin. the spin keeps a superposition that looks fine and carries a slightly wrong phase, and nobody finds out until tomography. the longer a batch runs, the more of that damage the next herald inherits.

with 15 attempts, a herald lands on average after seven earlier attempts in its batch, and that costs 1.92% of XX and YY. with 126 it costs about 15%. the paper measured this trade directly, and the widget runs the paper’s own error budget against those measurements:

fifteen tries, then resetfidelity against heralds per second as the batch length changes, computed from the paper's own error budget (Table S4) and timing. rings are the paper's measured points (Fig. 2d, labelled with attempts per batch) and the headline result.model, sweeping attempts per batchmeasured, as built
jump to
heralds per second-
fidelity-
silent photons-error on XX and YY
fake heralds-error on every correlator
attempting-share of the clock

the batch length only moves you along the curve. the return path moves the curve. slide share of those heard up and it swings toward the top right, with more heralds per attempt and fewer silent measurements per herald at the same time. a phase gate helps on both counts, because the photons that now vanish into the cavity would come back. the paper also suggests heralding on photons transmitted through the cavity, and rotating between several pre-initialized memories so each can reset without pausing the ion.

the source slider is less friendly. better collection raises the rate and dilutes the fake heralds, but it also pushes more silent photons into the memory per attempt, so the best batch length gets shorter. and on the phase gate + better optics preset, per-pair housekeeping (Doppler cooling plus a nuclear-spin initialization, about 48 µs on my fit) eats nearly a third of the clock at 6 kHz. past a few kHz, better optics stop paying off one for one.

87.9%

the fidelity is reconstructed from 9,000 heralds, a thousand in each of nine measurement settings, and it comes out at 87.9 ± 0.7% against a perfect Bell state. the paper’s error budget predicts 87.4% from independently measured pieces, so this budget closes too. by subsystem, the ion costs 4.3 points, the memory 3.9, frequency conversion 2.9 and detection 1.9.

the biggest individual lines are not where i’d have guessed:

  • 1.92% from silent photons dephasing the memory
  • 1.85% from fake heralds, the conversion noise
  • 1.26% from ion temperature and lens aberrations, a catch-all
  • 1.00% from the cavity’s finite on/off contrast
  • 1.00% from the ²⁹Si nuclear spin flipping during a run
  • 0.87% from pumping light leaking past the modulators after a herald

the two biggest errors come from loss and noise in the photonic chain. neither qubit is responsible for them. in this link, loss costs fidelity as well as rate: improve the source and the share of fake heralds falls, improve the return path and the silent-photon error falls. the paper says none of the dominant errors is fundamental, and the budget backs that up, because narrower filtering for the conversion noise, better modulator extinction, better cavity contrast and a better return path each go after a named line. none of them is free, either.

what i’d push on

the paper itself is careful. both budgets close, the data and analysis code are public, and the supplement even discloses a bug in the sequence script that disabled one attempt per batch in some datasets, which the authors found while writing it up. most of what follows is about framing.

the comparison isn’t like for like. the headline sets an ion-diamond link next to ion-ion links, and a computer needs ion-ion. through the memory that means two links and a swap, and swapping two 87.9% pairs leaves about 78%, whether you assume depolarizing noise or start from the paper’s measured correlators. so i put both architectures through the same back-of-envelope distillation: recurrence protocols with the best local rotations at every round, perfect local gates, as many memory qubits as needed, and no decay while pairs wait. ion-ion through the memory reaches 99% at about 47 pairs a second. the Duke two-photon link, treating its 93.7% lower bound as depolarizing noise, reaches 99% at about 53. at a fidelity a computer can use, the two architectures land in the same place. the estimate is generous to both, and an ion-ion experiment through the memory would settle it.

the rate is at the speed of computation and the fidelity isn’t. 12% infidelity against the 0.1% that fault-tolerance analyses assume is a factor of about a hundred. the same estimate takes this ion-diamond link to 99.9% in four rounds of distillation, at about 22 raw pairs per good one, which brings 1,032 a second down to under 50. that’s still a fine number for a first heterogeneous link, but it’s the one to hold up against 10 to 100 kHz.

fifteen seconds. at the wall-clock rate of 612 pairs a second, readout included, the 9,000 heralds behind the headline come from about 15 seconds of running. when the ion falls out of the trap or the diamond changes charge state, the authors stop, restore the system and throw away the whole dataset, and each run starts with a 30 s warm-up that brings the Pockels cell to thermal equilibrium. none of that is wrong. but the Bristol network post covered a network that ran for 157 hours, and an operator reads a rate as an average over a month. how often the ion leaves and the diamond ionizes is the number that turns 1 kHz into an availability figure.

the press release dropped the fidelity. IonQ’s announcement gives the rate and “more than four times faster than the previous record,” and no fidelity at all. the paper gives both, along with the 10 to 100 kHz target. read the paper.

where this goes next

most of what would move these numbers is already in print, some of it from this team and some from groups building rival designs.

1. ion-ion through the diamond. this is the experiment that turns the headline into the result. the pieces exist on the diamond side: a nuclear-spin register that stores entanglement while new links are made, Bell measurements across several photons in the store-and-forward work, and two diamond nodes entangled over deployed Boston fiber. the number to report is rate and fidelity together, after the swap.

2. a phase gate. Duan and Kimble proposed it in 2004: reflect the photon in both spin states and put the information in a spin-dependent phase, which works with strongly overcoupled cavities. it removes the factor-of-two ceiling on the bounce and most of the silent photons. the chip paper reports devices with cooperativity near 100 and about 85% fiber coupling, which is the hardware it needs.

3. a shorter clock. the controller and its cables take 346 of the 896 ns. an IonQ theory paper last year proposed skipping optical pumping altogether, which removes most of another 192, and on the stop-and-wait widget those two changes together are worth about 1.6×. after that it’s multiplexing, which ion groups already do: a Tsinghua node sends up to 44 time-bin modes and shuttles ions through the collection spot, so several attempts are in flight before the first herald comes back.

4. fidelity before the next rate record. either ion-memory pairs above 99% at kilohertz rates, or a distillation scheme printed with its rate cost. the estimate above puts the cost of 99.9% at a factor of about twenty today.

5. out of the dilution refrigerator. strained silicon vacancies have run at 1.5 K, which compact closed-cycle cryostats reach, and an MIT group is building tin-vacancy nodes on tunable photonic chips. conversion and interferometers belong on chips too, and thin-film lithium niobate at these wavelengths is the obvious platform.

6. benchmark against the other ways to pay once. this year brought three alternatives. Weizmann analyzed a single atom in a high-finesse cavity as the same kind of asynchronous memory interconnect, with nanosecond operations. MIT and Lincoln Lab argued for direct single-photon heralding between ions with integrated photonics and distillation. and USTC entangled an ytterbium ion with a rare-earth crystal memory across 75 m at 89% fidelity. a common figure of merit would help: Bell pairs per second per channel, at a stated fidelity, with housekeeping included, per square centimetre of chip and per watt of cryostat.

the short version

  • a modular quantum computer needs entangled pairs between modules at something like 10 to 100 kHz. the best photonic link between two ions makes 250 a second.
  • a two-photon herald needs both photons to arrive in the same attempt, so it pays the ion’s loss twice. here one photon from the ion bounces off a diamond memory and a single click heralds the pair.
  • the photon changes color from 493.5 to 737.45 nm, moves from polarization into time bins, bounces off the cavity, and comes back through the same interferometer, so the interferometer’s phase cancels and nothing needs stabilizing.
  • result: 1,032 pairs a second at 87.9% between a barium ion and a silicon vacancy, 612 a second with readout. both the rate budget and the error budget close.
  • the link spends 29.7 dB from clock to click, half of it on getting the photon into a fiber. the interferometer costs as much as the memory, and the amplitude gate can never return more than about half of the photons that reach it.
  • in decibels, the memory swaps the ion’s loss plus 3 dB for 13.3 dB of plumbing. it wins while the ion side costs more than about 9.4 dB, or 12.4 dB once you count the two links an ion-ion pair needs.
  • an attempt is 896 ns and the ion is busy for under a quarter of it. the link is stop-and-wait with a window of one, so 100 m of fiber halves the rate and a kilometre takes it to about 30 a second.
  • photons that reach the memory and go unheard dephase it, so longer batches buy rate with fidelity. the return path is the lever that improves both.
  • the two biggest errors come from the photonic chain: silent photons at 1.92% and conversion noise at 1.85%.
  • by my estimate, ion-ion through the memory and the Duke two-photon link both distill to 99% at about 50 pairs a second. the experiment that would settle it hasn’t been done.

in November 1977 a van driving around the San Francisco Bay Area sent packets through a radio network, the ARPANET and a satellite link, and they arrived. nobody remembers the throughput. what that demonstration showed was that three networks with nothing in common could carry one conversation through gateways, which is where the word internet comes from. this paper is the first quantum version of that i’ve seen running at a rate anyone could plan around. the photon that reaches the diamond has a different color and a different encoding from the one the ion emitted, and the only thing that made the whole trip is the qubit. somebody will probably beat the four-fold record soon, quite possibly the ion groups themselves. the gateway is the part i expect to last.